SKILL: Wan Prompt Engineering Master

Hermes Agent — Wan 2.2 (local) · 2.7 Thinking · 3.0 / 3.0 Prime (new)

DESCRIPTION

Prompting control for the Wan family. Three dialects share a subject, but tokenizers, audio defaults, and flags differ:

• 2.2 local (UMT5-XXL, Comfy native / WanVideoWrapper): cinematic English paragraph, motion-first on I2V, dual high-noise / low-noise MoE experts. Apache weights. This is what you run on 3090s.
• 2.7 Thinking Mode: same family, plus a plan-before-denoise pass. Structured sequential language (first / then). ~15 s class. Speed penalty ~30–60%.
• 3.0 / 3.0 Prime (the new Wan, hosted, public Aug 2026): up to 30 s, 20 references, native audio. Alibaba formula: subject+place+movement, then @Image1 jobs, numbered shots with [0-6s], voice / SFX / music as three layers. Prompt expansion and thinking are API flags, not prompt poetry. No 3.0 open drop — last Apache flagship is 2.2.

Never assume 2.2 negatives work at CFG 1 (Lightning). Never assume 3.0 stays silent if you omit sound. Never paste H3 <d> / At 00:03.500 into Wan 3.0.

Camera nodes / FLF / Animate → wan_camera_motion_control. 3.0 soundtrack → wan_audio_direction. @ refs / docs → wan_reference_control.

WAN 2.2 — LOCAL COMFY (MoE)